Research map / Large language models / Retrieval and knowledge
Are LLMs Really Not Knowledgeable? Mining the Submerged Knowledge in LLMs' Memory
Xingjian Tao, Yiwei Wang, Yujun Cai, Zhicheng Yang, Jing Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| measure Hits@5 instead of Hits@1 (top-k logit retention) (ablation) | LLaMA3-8b · DBpedia | Hits@k: 17.2 -> 57.9 | Improved |
| increase k to 50 (ablation) | LLaMA3-8b · DBpedia head | Hits@k: 48.3 (k=5) -> 83.4 (k=50) -> 90.5 (k=100) | Improved |
| scale LLaMA3 from 8b to 70b (ablation) | LLaMA3 · DBpedia head, k=100 | Hits@100: 90.5 -> 92.1 | No clear effect |
| scale LLaMA3 from 8b to 70b (ablation) | LLaMA3 · GoodReads head, k=100 | Hits@100: 67.8 -> 44.2 | Hurt |
| newer model generation (LLaMA3 vs LLaMA2) | LLaMA · DBpedia head, k=100 | Hits@100: 70.5 (LLaMA2-70b) -> 90.5 (LLaMA3-8b) | Improved |
| filter unsure-related tokens during decoding (ablation) | LLaMA3-70b · DBpedia head | accuracy (%): 11.2 -> 23.0 | Improved |
| filter unsure-related tokens during decoding (ablation) | Mistral-7b · DBpedia head | accuracy (%): 16.5 -> 16.7 | No clear effect |
| filter unsure-related tokens during decoding (ablation) | LLaMA2-70b · IMDB head | accuracy (%): 25.2 -> 25.2 | No clear effect |
| filter unsure-related tokens during decoding (ablation) | LLaMA3-8b · IMDB tail | accuracy (%): 2.2 -> 2.7 | No clear effect |
| open-domain vs domain-specific data | LLaMA3-8b · Hits@100 DBpedia vs IMDB head | Hits@100: 90.5 (DBpedia) vs 69.7 (IMDB) | Mixed |
Closest papers by meaning
- Do LLMs Really Forget? Evaluating Unlearning with Knowledge Correlation and Confidence Awareness (NeurIPS 2025 · similarity 0.90)
- Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs (ICLR 2026 · similarity 0.89)
- AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions (NeurIPS 2025 · similarity 0.89)
- PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbations (NeurIPS 2024 · similarity 0.89)
- Utilize the Flow Before Stepping into the Same River Twice: Certainty Represented Knowledge Flow for Refusal-Aware Instruction Tuning (AAAI 2025 · similarity 0.89)
- KScope: A Framework for Characterizing the Knowledge Status of Language Models (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
